Re: dcache shrink list corruption?

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sat, May 03, 2014 at 11:07:57AM -0700, Linus Torvalds wrote:
> Sure, umount itself should be serialized by the sb lock, so there
> should be only one umount dentry collector. But why wouldn't there be
> shrinkers active due to memory pressure?
> 
> generic_unmount_super() is called by ->kill_sb(), which is done
> *before* the superblock shrinker is unregistered  So any memory
> pressure during that will cause dentries to be shrunk other ways.
> 
> What am I missing?

This:
        if (!grab_super_passive(sb))
                return SHRINK_STOP;
before calling prune_dcache_sb().  grab_super_passive() returns with
->s_umount held shared on success (with down_read_trylock()) and ->kill_sb()
is called only with ->s_umount held exclusive.